NVIDIA DGX Spark: Local AI Hardware and Integration Guide
DGX Spark packages CPU, GPU, memory and networking into a compact system. Evaluate the complete platform and its software environment rather than treating it as a small interchangeable GPU server.
Hardware reference
The adapter rating describes the supplied power unit, not a measured constant wall draw. Confirm the delivered storage and accessory configuration against the quotation.
| Component | Published specification |
|---|---|
| CPU | 20-core Arm processor |
| GPU architecture | NVIDIA Blackwell |
| Memory | 128 GB unified LPDDR5x; 273 GB/s bandwidth |
| Storage | 1 TB or 4 TB NVMe M.2, depending on configuration |
| Networking | 10 GbE and ConnectX-7 connectivity |
| Power adapter | 240 W supplied adapter |
A sensible starting point
Shortlist the system for local development, model evaluation and workflows that benefit from an integrated NVIDIA environment. Decide which model artifacts, containers and tools you need before buying. The Arm host architecture matters when a dependency, binary extension or management tool assumes an x86 host.
For a shared service, define the request lengths, concurrency and latency requirement. A development experiment that works for one person is not automatically sized for a production team.
Read unified memory correctly
The installed memory serves the system as a whole. Model weights, execution state and other processes compete for that capacity. A model-size headline cannot specify the useful context length and user count for every runtime.
Do not infer a dual-system result simply by adding the memory labels. Distributed execution needs a supported software path and brings communication costs. If your plan requires two systems, validate that exact topology and record the assumptions.
Compare it with alternatives
A Mac Studio offers a different software ecosystem; a discrete-GPU workstation offers a different expansion path. Compare your chosen model, required dependencies, peak memory and representative workload output on each option. Treat published arithmetic throughput as one specification rather than an answer to the whole purchase decision.
Check the delivery package
Request the exact product configuration, storage, power adapter and cables, warranty provider and supported return route. Confirm the destination and support arrangements for the unit offered. Send your intended model and tools with the request so software requirements are considered alongside hardware.
Sources and documentation
- DGX Spark hardware overview ↗NVIDIA · checked 17 Nov 2026
Component Specifications; Power and Thermal Management